Mnemosyne OS · XPACEGEMS LLC
Sovereign memory, on machines you control.
Mnemosyne OS runs on your own hardware. Your vaults stay there, with AES-256 encryption at rest you arm under a key only you hold. No cloud required, ever. Built for enterprises, research labs, public institutions, and the teams building tomorrow’s machines.
Local-first · Optional AES-256 at rest · 77.1% LongMemEval-M · The human governs
Where you come from
Enterprise
Years of documents and conversations, searchable, on your own hardware. You decide what is kept.
Available today 02Research labs
77.1% on LongMemEval-M, every verdict published. And an open offer: a memory engine for your discipline, built on your corpus.
Open collaboration 03Public sector
Sovereignty as the default: your jurisdiction, your hardware, no mandatory foreign cloud.
Open to discuss 04Banking & finance
A client’s full context in one place, on the bank’s own infrastructure. No compliance certification yet.
Pilot wanted 05Robotics, the frontier
A robot cannot outsource its memory. Latency breaks reflexes, and a lost connection cannot mean amnesia. On board, encrypted, offline. No robot runs this today.
Direction, not product Not an organization Individuals, freelancers and students: the product lives on the product site, free for personal use. mnemosyne-os.io →Each tile says where it really stands. Available today means deployed and running. Pilot wanted means the engine exists and the deployment does not. Direction means we have thought it through and built nothing. Better you read it here than find out on a call.
A number that arrives with its paperwork.
Our retrieval engine scores 77.1% on LongMemEval-M, a public benchmark for long-term conversational memory, under the strict judge. What matters is how the number was obtained. Every question gets its own verdict, the run was replayed and matched verdict for verdict, and the grading protocol ships with the score.
The grader, the verdicts and the method are published. You can redo the math without taking our word for it. The engine stays private: that is where reproducible measurement stops and industrial property starts. What we missed is documented next to what we scored.
LongMemEval-M · 77.1% strict · one verdict per question · replayed, verdict for verdict · misses documented
The benchmark itself is public and third-party: LongMemEval on GitHub →
The full dossier (numbers, protocol, receipts) lives on the product site: mnemosyne-os.io/benchmark →
Mnemosyne Labs
The first work is deposited. A technical whitepaper on the architecture, and the audit kit behind the benchmark, both under permanent identifiers and attached to an ORCID. The lab notes say what was measured, and what turned out to be wrong.
Published Technical whitepaper, DOI 10.5281/zenodo.21728284Benchmark audit kit, DOI 10.5281/zenodo.21727140 ORCID 0009-0009-1087-3917
References The research lineage we build on
Two pieces of public research sit under the architecture. One we extend, one we measure ourselves against.
-
The “LLM as an operating system” idea: a virtual context managed the way an OS manages memory. Mnemosyne OS goes further. Not a library you import per session, but a daemon that runs continuously, isolates agents and persists.
arXiv:2310.08560 · MemGPT · Packer et al. · UC Berkeley · 2023
-
The yardstick we submit to: a third-party academic benchmark for long-term memory, not a metric of our own making. Our 77.1% is measured on it.
arXiv:2410.10813 · LongMemEval · Wu et al. · ICLR 2025
Contact
Anything that fits no section above. Every form on this page reaches a human, this one too.